Research Question

• Can we build an early warning system of food security in areas where data is scarce and data collection is costly ? (Hutchinson,1991)
  • Timely and accurate targeting is essential for aid and humanitarian responses
• How to make use of publicly available and economically meaningful data to fill the data gap?
  • Price data of the main agricultural markets are collected monthly or weekly
  • Precipitation/temperature/soil quality from remote sensing are relevant to agricultural production
  • Distance to roads and markets reflects access to market and information
• Use supervised learning approach that to achieve higher predicative power and remain interpretable.

Preview of Results

• Out-of-sample predictions from our model explains up to 50%-67% of cluster level variations in Malawi and up to 76% in Tanzania for HDDS and FCS measures.
  • 0.4 r² on consumption expenditures and 0.6 r² on asset index (Jean et al., 2016)
  • 0.76 r² on WI and 0.27 r² on income in Bangladesh (Steele et al,2017)

• Using the same sets of features, a machine learning model outperforms a baseline linear model by 20 - 50% in high dimensional settings.

• Decreased the proportion of estimates that overestimate Food Security outcomes

• Validates the “A Prototype for Predicting Food Insecurity Using Readily Available Data” paper with Tanzania.

Literature Review

• Night lights data (Chen and Nordhaus 2011;Henderson et al. 2012.) does a good job at predicating economic activity but the variation is little in areas of the extreme poor or in urban areas.

• Mobile phone data (Blumenstock et al.,2015; Steele et al.,2017) is complimentary to remote sensing data but expensive.

• Very High resolution satellite imagery are becoming cheaper but highly unstructured and contains measurement error (Engstrom, 2018; Donaldson and Storeygard, 2016).

• Convolutional Neural Network (CNN) models (Jean et al., 2016; Babenko et al. 2017) explains average 46% of the variation at village level but requires an enormous amount of training data and are computationally extensive. Interpret-ability and repeat-ability are less promising.

Framework

• Understanding of food security. The definition of food security has various characteristics. The use of metrics can lead to quite divergent rankings of the same population (Steele et al. 2017)

• Geo-referenced household surveys (LSMS data) allow us to explore the spatial-temporal variations in food security measures.

• The sampling framework allows us to observe a nationally representative HH sample in different months and agroecological zones.

• Explain these variations by the spatial-temporal variation in food availability and in food access.
  • Align weather data to crop growing season
  • Align households to the most relevant market price

• Utilize the interaction and higher order terms of the variables

Indicators to measure food security

• Dietary diversity and food frequency

  • Food Consumption Score (FCS)

  • Household Dietary Diversity Scale ( HDDS )

  • Captures quality and diversity

• Consumption behaviors

  • Reduced Coping Strategy Index (rCSI)

  • Capture the element of quantity or sufficiency

Indicators to measure food security

correlation between food security measures

Source: How Do Different Indicators of Household Food Security Compare? Empirical Evidence from Tigray by Maxwell, Coates and Vaitla 2013.

Temporal Variation

Graph of Food Security by Month Malawi

Spatial Variation

Map of Food Security Maps in Malawi by month

Spatio-temporal variation

Food Security Maps in Tanzania

Predication

• Split the entire sample into training and testing
• Use Malawi 2010 to predict Malawi 2013; and a randomly split training and testing sample for Tanzania

Importance out of sample predication

Overfit

Avoid over-fit by using simple models, cross-validation, regularization, ensemble learning

Over vs under fit

Models

• Set of models deal with high dimensional problem with different types regularization

• Maximum Likelihood plus a constraint ( same interpretation )

  • Lasso (L1 norm, prefers smaller model)
  • Ridge (L2 norm, prefers large model)
  • Elastic Net (Combination of the two)
• Ensemble leaning models that are great with positively skewed and high dimensional data:
  • Random Forest
  • Gradient boosting
  

  Random Forest

Data

• Price data (maize and rice price and market thinness at the nearest district market)

• Weather

  • Total rainfall in growing season (October to April)
  • Maximum of days without rain
  • First day of rain
  • Mean temperature
  • Growing degree days (days that average temperature between 5-32 Degree Celsius) (Deschenes and Greenstone 2007)

• Soil quality: slope, elevation, soil nutrition

• Roof/floor type, household asset index

• Mobile phone ownership: access to financial resources, remittance flow and economic (Eagle et al 2010, Blumenstock et al 2016)

Main Results for Malawi

Scatterplots for Malawi (FCS)

Scatterplots of different models

Scatterplots for Malawi (HDDS)

Scatterplots of different models

Main Results for Tanzania

Scatterplots for Tanzania (FCS)

Scatterplots of different models

Scatterplots for Tanzania (HDDS)

Scatterplots of different models

Future Steps

• Vary the time gap between training and testing ( train and test on the a subset of data that are only several weeks/month apart)
• Trained on a pooled data set across different countries V.S. Fit models on each individual country with the same procedure
• Predict “now”: countries/areas that are not surveyed and suggest areas that are likely to have a food shortage.

Limitations

• limited to gradual food insecure cases and can’t predict sudden, abrupt threat to food security (natural disaster, war and conflict)